Design-partner program · seed stage

Prevention, not detection. The financial control plane for AI agents.

Tarmac sits between your agents and the compute they consume, holds the budget, and enforces it before money is spent. Per-run, monthly, and yearly caps as hard limits — atomic authorization holds, modeled on a card network. No more month-end surprises.

412 on the runway No card · No spam SOC 2 Type I — Q2
You're on the list. We'll reach out from crew@tarmac.dev.
// Live authorization ledger Live
Monthly budget $184,200 / $250,000 11 days left · pace 92%
$0$125K$250K · cap
support-triageP0$0.34admit · L
revenue-opsP0$1.92admit · XL
code-refactorP2$4.10$1.20throttle · M
data-pipeline-euP1$2.65admit · L
research-crawlerP3$3.85hold · over-cap
eval-harnessP2$0.92admit · M
admit throttle hold 2,141 holds/min · 0 breaches
// 01 · The problem

Autonomous agents now spend money without a real-time check. Finance sees the bill after it lands.

Coding, research, monitoring and data-pipeline agents already hold real budgets. Today's tools observe, alert and throttle — none of them prevent a breach. The pain is industry-wide and compounding faster than any adjacent vendor is moving to govern it.

// Median enterprise · monthly LLM bill · YoY into 2026
7.2×

Year-over-year growth in the median enterprise's monthly LLM spend — outpacing every other line item in the cloud bill, with no governance layer in front of it.

FinOps Foundation · 2026 State of FinOps for AI

Cost compounds with autonomy.

A single multi-step agentic decision cycle can cost three orders of magnitude more than a simple call. Every new tool, every retry, multiplies the bill.

Simple LLM call~$0.001
Tool-augmented response$0.02–0.08
Multi-step agentic decision cycle$0.10–$1.00
// 02 · How it works

Tarmac admits, decides, enforces, settles — like a card network for compute.

Every run is gated through an atomic authorization hold against the per-run, monthly, and yearly caps. If it doesn't fit, it doesn't run. If it fits, Tarmac picks the heaviest model the live price and remaining budget allow.

01 / Onboard

Set budgets and priority.

Define a per-run max, monthly and yearly caps. Each agent carries a description and a priority tier — revenue-critical work is protected over nice-to-have work.

caps · tiers
02 / Watch

Track live price & burn.

The Compute Price Signal continuously aggregates model and GPU pricing. The ledger keeps remaining budget to the cent, across every concurrent run.

price · burn
03 / Decide

Pick the model tier.

For every run, Tarmac selects the heaviest model that the live price and remaining budget allow — lighter to protect the budget, heavier when there's room.

tier · S→XL
04 / Enforce

Apply, hold, or drop.

An atomic authorization hold is placed before the call is admitted. Over-cap runs are throttled to a lighter tier — or held entirely if no tier fits.

admit · hold
05 / Report

Attribute every dollar.

One reconciled ledger powers a live ops view and a finance-grade set of books: statements, itemized receipts, attribution and burn-rate projection.

books · close
// 03 · Two reconciled surfaces

One ledger. Two views. Built for two buyers at once.

The platform leader gets a live control room. The CFO gets a managed budget with statements and forecast. Both views read from the same enforced ledger — never out of sync, never reconciled by hand.

// Surface 01 For platform & AI-infra leaders

Agent Operations Dashboard.

A live control room over the entire fleet. Watch admits and holds in real time, rebalance by priority before a cap is breached, and ship without fear of the bill.

  • Fleet view by agent, priority tier, and laneEvery agent's burn, hold rate, and active tier at a glance.
  • Priority-aware automatic rebalancingTighten budgets shed P3 work first; P0 keeps full tier.
  • Per-agent rules & soft alertsCatch drift before it becomes a hold spike.
// Surface 02 For the CFO & finance

Financial Console.

AI compute spend that finally behaves like a managed budget. Monthly statements, itemized receipts, clean attribution to team and agent, and a burn-rate projection that doesn't move at month-end.

  • Period statements with full attributionBy cost center, team, agent, and lane — close the books on AI cost.
  • Burn-rate projection to month-endForecast updates with every authorization, not at month-end.
  • Audit-ready ledger with itemized receiptsEvery run, every hold, every dollar — reconciled to the cent.
// 04 · The core mechanic

Pin the lane. Move the model. Within the budget, always the heaviest tier that fits.

Each agent's compute lane — an inference-API model family or a fixed GPU type — is set once at onboarding. What Tarmac moves dynamically, per run, is the model tier the agent runs: lighter to protect the budget, heavier when there's room.

One lane per agent.

Onboard-time decision: family, region, fallback. Predictable for security review, simple for compliance, frozen against drift.

revenue-opspriority · P0
S
M
L
XL · heaviest fit
data-pipelinepriority · P1
S
M
L · fits the cap
XL
code-refactorpriority · P2
S
M · throttled
L
XL
researchpriority · P3
S · floor
M
L
XL
// lighter ← model tier → heavierrecalc · every authorization
// 05 · The wedge

Observability sees. Gateways throttle. Only Tarmac enforces.

Adjacent tools exist, and major platforms are moving in — but they cluster into five groups, and the same gap runs through all of them. None enforce a hard cap before spend. None rebalance a fleet by business priority. None produce books that engineering and finance both trust.

Category
Pre-spend hard cap
Priority-aware rebalance
Finance-grade books
AI gatewaysPortkey · LiteLLM · Kong AI
throttle & alert
per-key quota
developer-grade
FinOps / cloud-costCloudZero · Vantage · Finout
retrospective
not in path
allocation only
AI observabilityHelicone · Langfuse · Datadog
traces, not control
monitoring
not the system of record
Model routersMartian · Not Diamond · OpenRouter
per-request only
per-request only
no ledger
Platform-bundledDatabricks · hyperscalers
single-platform
shallow
locked-in
Tarmac// the financial control plane
atomic authorization holds
by business priority
one reconciled ledger
// 06 · Pricing

3% of what we govern. 1% of what we save. One model. No tiers.

Tarmac never resells compute. The governance fee scales with the spend we protect; the performance fee earns only when we deliver savings. Our incentives point the same direction as yours.

// Governance fee
3%of governed AI compute spend · annual
Minimum · $12,000 / year

A proportion of the spend Tarmac protects — never a flat platform tax. Same rate from your first agent to your thousandth.

  • Authorization-hold engine · per-run, monthly, yearly caps
  • Priority-aware fleet rebalancing under budget pressure
  • Agent Operations Dashboard · live fleet view
  • Financial Console · statements, exports, audit ledger
  • Unlimited users · unlimited agents · SOC 2 Type II
// Performance fee
1%of realized savings · quarterly in arrears
Billed only on savings actually delivered

For every run, the gap between what the task would have cost at the heaviest model and what it actually cost. Summed across the fleet, traceable to every receipt.

  • Tarmac earns more only when it has saved you more
  • Savings reporting on every receipt, statement and rollup
  • No charge in any quarter Tarmac doesn't deliver savings
  • Never a percentage of your cloud bill or provider invoice
  • Never a markup on compute — Tarmac is not a reseller
Annual price = max($12,000, 3% × governed spend) + 1% × realized savings. · Compute runs in your own accounts. · See worked examples & estimator →
// 07 · Built for the security review

Proxy-first. Fail-closed. An in-path vendor is only a vendor if the CISO signs off — so we built for the review.

01 / Proxy-first integration

No long-lived customer credentials.

Tarmac sits in the request path as a proxy with short-lived, scoped tokens issued per workspace. No standing read on your model provider account. No keys at rest in our control plane.

scoped-tokens · per-workspace
02 / Fail-closed by default

If Tarmac can't authorize, the run doesn't happen.

The ledger is the system of record. When the budget engine is unhealthy, the proxy refuses to admit rather than letting a run leak through. Budget breach is a designed-against, tested property.

concurrency-invariant · CI gate
03 / SOC 2 from day one

Type I in Q2. Type II observation window opens Q3.

GRC platform live before the first design partner. Annual third-party penetration tests, a public trust center, and SOC 2 Type II on the calendar before the first enterprise opportunity — not after.

SOC 2 · vCISO · pentest · trust center
// Compliance roadmap
✓ GRC platform live SOC 2 Type I · Q2 2026 SOC 2 Type II · Q4 2026 First pen test · Q2 2026 ISO/IEC 27001 · 2028
// 08 · Design partner program

Get the first ledger.

We're working with a small cohort of AI-native and digital-native teams to put Tarmac into production on real agent fleets. Discounted pricing, direct access to the founder, and the right to shape the product before it's set.

// What you get

Six pilots in Year 1. Two are already spoken for.

  • Pilot-pricing: 50% off the first 12 months
  • Direct line to the founder & engineering
  • Public reference logo & one published case study
  • MVP onboarding in Q3 2026, Phase 1 product
  • Right of first refusal on the Series A
// 09 · FAQ